We provide two pre-trained models: human to anime and human to cats
Run the following command to translate the demo inputs. Note that this command works the best for models trained using twingan. You'll have to change this a little to make it work with image_generation models.
# Make sure you're under
python inference/image_translation_infer.py \
--model_path="/PATH/TO/MODEL/"
--image_hw=256 # 256 for anime, 128 for cats.
--input_tensor_name="sources_ph"
--output_tensor_name="custom_generated_t_style_source:0"
--input_image_path="./demo/inference_input/cropped"
--output_image_path="./demo/inference_output/temp"
You can modify your training script into eval script by changing the following flags. (If you are using pggan_runner, it automatically sets the checkpoint_path and the eval_dir for you.)
--checkpoint_path="/PATH/TO/CHECKPOINT"
--eval_dir="/PATH/TO/SAVE/EVAL/RESULT"
--dataset_split_name=validation
--is_training=False
--do_custom_eval=True
--calc_swd=True
--use_tf_swd=False
--swd_num_images=8192
--swd_save_images=False
Here's an example script that evaluates the PGGAN output using msss.
python image_generation.py
--train_dir="/PATH/TO/CHECKPOINT"
--checkpoint_path="/PATH/TO/CHECKPOINT"
--eval_dir="/PATH/TO/SAVE/EVAL/RESULT"
--batch_size=32
--dataset_name="image_only"
--dataset_dir="/PATH/TO/DATASET"
--dataset_split_name=validation
--is_training=False
--train_image_size=32
--preprocessing_name="danbooru"
--generator_network="pggan"
--loss_architecture=dragan
--gradient_penalty_lambda=0.25
--do_custom_eval=True
--calc_swd=True
--use_tf_swd=False
--swd_num_images=8192
--swd_save_images=False
If you see ModuleNotFoundError: No module named 'util_io', you're probably not under the TwinGAN directory, or you have not set PYTHONPATH correctly. See Issue 10 and 3 for details.